English Vocabulary for MCP (Model Context Protocol) Developers

The Model Context Protocol (MCP) is rapidly becoming the standard way to connect AI assistants to external data and tools. If you work with MCP or plan to, knowing the precise English vocabulary used by developers on GitHub, in documentation, and in team discussions is essential for professional communication.

Key Vocabulary

Host In MCP, the host is the application that runs the AI model and manages one or more MCP clients. Think of it as the orchestrator — it decides which MCP servers to connect to and controls what the model is allowed to do. Example: “The IDE acts as the MCP host and spawns a separate client connection for each configured server.”

Client A client is a component inside the host that maintains a one-to-one connection with a single MCP server. Developers often say a client “speaks to” or “connects to” a server. Example: “Each MCP client handles the lifecycle of its server connection, including initialization and shutdown.”

Server An MCP server is a lightweight process that exposes capabilities — tools, resources, or prompts — to the host via the protocol. Servers are usually purpose-built: one for a database, one for a file system, one for a web search API. Example: “We built a custom MCP server that exposes our internal ticketing system as a set of tools.”

Resource A resource is a data source that a server makes available for the model to read. Resources have URIs and can represent files, database rows, API responses, or any structured content. Developers say they “expose” or “serve” resources. Example: “The server exposes a resource at file:///workspace/config.yaml so the model can read the current configuration.”

Tool A tool is a callable function that the model can invoke to take an action or fetch data. Unlike resources (which are read-only), tools can have side effects. The model “calls” or “invokes” a tool; the server “registers” or “exposes” it. Example: “We registered a create_ticket tool on the MCP server so the assistant can open Jira issues automatically.”

Prompt In MCP, a prompt is a pre-defined template that the server offers to the host. Prompts help users start common workflows with a structured input. The server “provides” prompts; the host “lists” and “renders” them. Example: “The server provides a code_review prompt template that pre-fills the context with the current diff.”

Sampling Sampling is the mechanism by which an MCP server asks the host to run an LLM inference on its behalf. This lets servers build agentic loops without needing direct model access. Developers say a server “requests sampling” or “triggers a sampling call.” Example: “The server requests sampling to summarize a long document before returning a concise tool result.”

Transport Layer The transport layer defines how messages travel between the client and the server. The two main options are stdio (standard input/output, used for local processes) and SSE (Server-Sent Events, used for remote HTTP servers). Developers “configure” or “select” the transport. Example: “For local development we use the stdio transport; in production the server runs over SSE behind a reverse proxy.”

Common Phrases and Collocations

“expose a tool via MCP” Use this phrase when describing what a server makes available to the model. The verb “expose” is standard — avoid saying “publish” or “open.” Example: “This PR exposes three tools via MCP: one to query the database, one to write files, and one to send emails.”

“grant resource access” Used when discussing permissions — the host or user decides which resources a client is allowed to read. Example: “The user must explicitly grant resource access before the model can read files outside the project directory.”

“configure the MCP host” Refers to setting up the host application (often via a JSON config file) to know which servers to launch and how to connect. Example: “To configure the MCP host, add the server entry to your mcp_config.json with the command and arguments.”

“initialize the connection” The handshake phase when client and server exchange capabilities. Both sides “negotiate” or “agree on” supported protocol versions. Example: “During initialization the client and server negotiate the protocol version and exchange their capability lists.”

“context window management” The practice of controlling how much content is injected into the model’s context from MCP resources, so the window does not overflow. Example: “Good context window management means summarizing large resources before passing them to the model.”

Practical Sentences to Practice

1. “I’m going to expose our search index as a resource on the MCP server so the model can query it directly.”
2. “The client failed to initialize the connection because the server process exited before the handshake completed.”
3. “We switched from stdio to SSE transport so multiple users can share the same remote MCP server.”
4. “Make sure the host grants resource access only to the directories the user has opened in the workspace.”
5. “The sampling request timed out — I think the model inference is taking too long inside the agentic loop.”

Common Mistakes to Avoid

Confusing “client” and “server” In everyday networking we often think of the server as the powerful machine. In MCP, the server is actually the smaller helper process. The host (with the AI model) is the powerful side. Always clarify: “the MCP server exposes tools” and “the MCP client connects to the server.”

Using “publish” instead of “expose” MCP documentation consistently uses “expose” for tools and resources. Saying “the server publishes a tool” sounds unusual in this context. Stick with “expose” or “register.”

Saying “run sampling” instead of “request sampling” Sampling is requested by the server and executed by the host. The server does not run inference itself — it requests that the host do so. Use “the server requests sampling” to be precise.

Summary

MCP introduces a precise vocabulary that distinguishes hosts, clients, servers, resources, tools, prompts, and transport layers. Each term has a specific role in the protocol, so using them correctly signals technical fluency in AI engineering discussions. Practice by reading the official MCP specification and open-source MCP server repositories, paying attention to how contributors phrase issues and pull request descriptions. The more you read real developer communication in this space, the faster these terms will become natural.